qZACH-ViT: Quantization-Aware Intrinsic Explanations with Recursive Attribution-Stabilized Optimization

Short summary

qZACH-ViT introduces a quantization-aware, intrinsically explainable Vision Transformer for medical image classification, combining INT8 mixed-precision deployment with recursive attribution-stabilized optimization (RASO). Across seven MedMNIST datasets with 280 runs, the INT8 model matches or exceeds FP32 baselines while achieving 99.97% prediction agreement and 70% smaller ONNX artifacts with up to 2.39x CPU speedup. RASO improves attribution stability and sufficiency error, making the model both deployable and interpretable for clinical use.

• •Quantization-aware intrinsically explainable ViT for medical imaging with INT8 ONNX deployment
• •RASO optimization stabilizes attribution gradients; 70% model size reduction with 2.39x CPU speedup
• •99.97% prediction agreement between FP32 and INT8 across 964,920 comparisons on 7 MedMNIST datasets

Generated with AI, which can make mistakes.